Proceedings of the Matlab DSP Conference 1999, Espoo, Finland, November 16–17, pp. 35–40, 1999.
Self-organizing map in Matlab: the SOM Toolbox
Juha Vesanto, Johan Himberg, Esa Alhoniemi and Juha Parhankangas
Laboratory of Computer and Information Science, Helsinki University of Technology, Finland
Abstract
The Self-Organizing Map (SOM) is a vector
quantization method which places the prototype vectors on
a regular low-dimensional grid in an ordered fashion.
This makes the SOM a powerful visualization tool. The
SOM Toolbox is an implementation of the SOM and its
visualization in the Matlab 5 computing environment. In
this article, the SOM Toolbox and its usage are shortly
presented. Also its performance in terms of computational
load is evaluated and compared to a corresponding C-
program.
1. General
This article presents the (second version of the) SOM
Toolbox, hereafter simply called the Toolbox, for Matlab
5 computing environment by MathWorks, Inc. The SOM
acronym stands for Self-Organizing Map (also called
Self-Organizing Feature Map or Kohonen map), a popular
neural network based on unsupervised learning [1]. The
Toolbox contains functions for creation, visualization and
analysis of Self-Organizing Maps. The Toolbox is
available free of charge under the GNU General Public
License from http://www.cis.hut.fi/projects/somtoolbox.
The Toolbox was born out of need for a good,
easy-to-use implementation of the SOM in Matlab for
research purposes. In particular, the researchers
responsible for the Toolbox work in the field of data
mining, and therefore the Toolbox is oriented towards that
direction in the form of powerful visualization functions.
However, also people doing other kinds of research using
SOM will probably find it useful — especially if they
have not yet made a SOM implementation of their own in
Matlab environment. Since much effort has been put to
make the Toolbox relatively easy to use, it can also be
used for educational purposes.
The Toolbox — the basic package together with
contributed functions — can be used to preprocess data,
initialize and train SOMs using a range of different kinds
of topologies, visualize SOMs in various ways, and
analyze the properties of the SOMs and data, e.g. SOM
quality, clusters on the map and correlations between
variables. With data mining in mind, the Toolbox and the
SOM in general is best suited for data understanding or
survey, although it can also be used for classification and
modeling.
2. Self-organizing map
A SOM consists of neurons organized on a regular low-
dimensional grid, see Figure 1. Each neuron is a d-
dimensional weight vector (prototype vector, codebook
vector) where d is equal to the dimension of the input
vectors. The neurons are connected to adjacent neurons by
a neighborhood relation, which dictates the topology, or
structure, of the map. In the Toolbox, topology is divided
to two factors: local lattice structure (hexagonal or
rectangular, see Figure 1) and global map shape (sheet,
cylinder or toroid).
2 2
1 1
0 0
Figure 1. Neighborhoods (0, 1 and 2) of the centermost
unit: hexagonal lattice on the left, rectangular on the right.
The innermost polygon corresponds to 0-, next to the 1-
and the outmost to the 2-neighborhood.
The SOM can be thought of as a net which is spread to
the data cloud. The SOM training algorithm moves the
weight vectors so that they span across the data cloud and
so that the map is organized: neighboring neurons on the
grid get similar weight vectors. Two variants of the SOM
training algorithm have been implemented in the Toolbox.
In the traditional sequential training, samples are
presented to the map one at a time, and the algorithm
gradually moves the weight vectors towards them, as
shown in Figure 2. In the batch training, the data set is
presented to the SOM as a whole, and the new weight
vectors are weighted averages of the data vectors. Both
algorithms are iterative, but the batch version is much
faster in Matlab since matrix operations can be utilized
efficiently.
For a more complete description of the SOM and its
implementation in Matlab, please refer to the book by
Kohonen [1], and to the SOM Toolbox documentation.
X with som_batchtrain. Of course, also increase in
BMU
Figure 2. Updating the best matching unit (BMU) and
its neighbors towards the input sample marked with x.
The solid and dashed lines correspond to situation before
and after updating, respectively.
3. Performance
The Toolbox can be downloaded for free from
http://www.cis.hut.fi/projects/somtoolbox. It requires no
other toolboxes, just the basic functions of Matlab (version
5.1 or later). The total diskspace required for the Toolbox
itself is less than 1 MB. The documentation takes a few
MBs more.
The performance tests were made in a machine with 3
GBs of memory and 8 250 MHz R10000 CPUs (one of
which was used by the test process) running IRIX 6.5
operating system. Some tests were also performed in a
workstation with a single 350 MHz Pentium II CPU, 128
MBs of memory and Linux operating system. The Matlab
version in both environments was 5.3.
The purpose of the performance tests was only to
evaluate the computational load of the algorithms. No
attempt was made to compare the quality of the resulting
mappings, primarily because there is no uniformly
recognized “correct” method to evaluate it. The tests were
performed with data sets and maps of different sizes, and
three training functions: som_batchtrain,
som_seqtrain and som_sompaktrain, the last of
which calls the C-program vsom to perform the actual
training. This program is part of the SOM_PAK [3],
which is a free software package implementing the SOM
algorithm in ANSI-C.
Some typical computing times are shown in Table 1. As
a general result, som_batchtrain was clearly the
fastest. In IRIX it was upto 20 times faster than
som_seqtrain and upto 8 times faster than
som_sompaktrain. Median values were 6 times and 3
times, respectively. The som_batchtrain was
especially faster with larger data sets, while with a small
set and large map it was actually slower. However, the
latter case is very atypical, and can thus be ignored. In
Linux, the smaller amount of memory clearly came into
play: the marginal between batch and other training
functions was halved.
The number of data samples clearly had a linear effect
on the computational load. On the other hand, the number
of map units seemed to have a quadratic effect, at least
input dimension increased the computing times: about
two- to threefold as input dimension increased from 10 to
50. The most suprising result of the performance test was
that especially with large data sets and maps, the
som_batchtrain outperformed the C-program (vsom
used by som_sompaktrain). The reason is probably
the fact that in SOM_PAK, distances between map units
on the grid are always calculated anew when needed. In
SOM Toolbox, all these are calculated beforehand.
Likewise for many other required matrices.
Indeed, the major deficiency of the SOM Toolbox, and
especially of batch training algorithm, is the expenditure
of memory. A rough lower bound estimate of the amount
of memory used by som_batchtrain is given by:
8(5(m+n)d + 3m2) bytes, where m is the number of
map units, n is the number of data samples and d is the
input space dimension. For [3000 x 10] data matrix and
300 map units the amount of memory required is still
moderate, in the order of 3.5 MBs. But for [30000 x 50]
data matrix and 3000 map units, the memory requirement
is more than 280 MBs, the majority of which comes from
the last term of the equation. The sequential algorithm is
less extreme requiring only one half or one third of this.
SOM_PAK requires much less memory, about 20 MBs for
the [30000 x 50] case, and can operate with buffered data.
Table 1. Typical computing times. Data set size is
given as [n x d] where n is the number of data samples
and d is the input dimension.
data size map units batch seq sompak
IRIX
[300x10] 30 0.2 s 3.1 s 0.9 s
[3000x10] 300 7s 54 s 17 s
[30000x10] 1000 5 min 19 min 9 min
[30000x50] 3000 27 min 5.7 h 75 min
Linux
[300x10] 30 0.3 s 2.7 s 1.9 s
[3000x10] 300 24 s 76 s 26 s
[30000x10] 1000 13 min 40 min 15 min
4. Use of SOM Toolbox
4.1. Data format
The kind of data that can be processed with the
Toolbox is so-called spreadsheet or table data. Each row
of the table is one data sample. The columns of the table
are the variables of the data set. The variables might be the
properties of an object, or a set of measurements measured
at a specific time. The important thing is that every sample
has the same set of variables. Some of the values may be
missing, but the majority should be there. The table
representation is a very common data format. If the
available data does not conform to these specifications, it
can usually be transformed so that it does.
The Toolbox can handle both numeric and categorial
data, but only the former is utilized in the SOM algorithm.
In the Toolbox, categorial data can be inserted into labels
associated with each data sample. They can be considered
as post-it notes attached to each sample. The user can
check on them later to see what was the meaning of some
specific sample, but the training algorithm ignores them.
Function som_autolabel can be used to handle
categorial variables. If the categorial variables need to be
utilized in training the SOM, they can be converted into
numerical variables using, e.g., mapping or 1-of-n
coding [4].
Note that for a variable to be “numeric”, the numeric
representation must be meaningful: values 1, 2 and 4
corresponding to objects A, B and C should really mean
that (in terms of this variable) B is between A and C, and
that the distance between B and A is smaller than the
distance between B and C. Identification numbers, error
codes, etc. rarely have such meaning, and they should be
handled as categorial data.
4.2. Construction of data sets
First, the data has to be brought into Matlab using, for
example, standard Matlab functions load and fscanf.
In addition, the Toolbox has function som_read_data
which can be used to read ASCII data files:
sD = som_read_data(‘data.txt’);
The data is usually put into a so-called data struct,
which is a Matlab struct defined in the Toolbox to group
information related to a data set. It has fields for numerical
data (.data), strings (.labels), as well as for
information about data set and the individual variables.
The Toolbox utilizes many other structs as well, for
example a map struct which holds all information related
to a SOM. A numerical matrix can be converted into a
data struct with: sD = som_data_struct(D). If the
data only consists of numerical values, it is not actually
necessary to use data structs at all. Most functions accept
numerical matrices as well. However, if there are
categorial variables, data structs has be used. The
categorial variables are converted to strings and put into
the .labels field of the data struct as a cell array of
strings.
4.3. Data preprocessing
Data preprocessing in general can be just about
anything: simple transformations or normalizations
performed on single variables, filters, calculation of new
variables from existing ones. In the Toolbox, only the first
of these is implemented as part of the package.
Specifically, the function som_normalize can be used
to perform linear and logarithmic scalings and histogram
equalizations of the numerical variables (the .data
field). There is also a graphical user interface tool for
preprocessing data, see Figure 3.
Scaling of variables is of special importance in the
Toolbox, since the SOM algorithm uses Euclidean metric
to measure distances between vectors. If one variable has
values in the range of [0,...,1000] and another in the range
of [0,...,1] the former will almost completely dominate the
map organization because of its greater impact on the
distances measured. Typically, one would want the
variables to be equally important. The standard way to
achieve this is to linearly scale all variables so that their
variances are equal to one.
One of the advantages of using data structs instead of
simple data matrices is that the structs retain information
of the normalizations in the field .comp_norm. Using
function som_denormalize one can reverse the
normalization to get the values in the original scale: sD =
som_denormalize(sD). Also, one can repeat the
exactly same normalizations to other data sets.
All normalizations are single-variable transformations.
One can make one kind of normalization to one variable,
and another type of normalization to another variable.
Also, multiple normalizations one after the other can be
made for each variable. For example, consider a data set
sD with three numerical variables. The user could do a
histogram equalization to the first variable, a logarithmic
scaling to the third variable, and finally a linear scaling to
unit variance to all three variables:
sD = som_normalize(sD,'histD',1);
sD = som_normalize(sD,'log',3);
sD = som_normalize(sD,'var',1:3);
The data does not necessarily have to be preprocessed
at all before creating a SOM using it. However, in most
real tasks preprocessing is important; perhaps even the
most important part of the whole process [4].
 4.4. Initialization and training
There are two initialization (random and linear) and
two training (sequential and batch) algorithms
implemented in the Toolbox. By default linear
initialization and batch training algorithm are used. The
simplest way to initialize and train a SOM is to use
function som_make which does both using automatically
selected parameters:

sM = som_make(sD);

The training is done is two phases: rough training with

large (initial) neighborhood radius and large (initial)
learning rate, and finetuning with small radius and
2 learning rate. If tighter control over the training
1
0
parameters is desired, the respective initialization and
−1 training functions, e.g. som_batchtrain, can be used
directly. There is also a graphical user interface tool for
initializing and training SOMs, see Figure 4.
4.5. Visualization and analysis
There are a variety of methods to visualize the SOM. In
the Toolbox, the basic tool is the function som_show. It
can be used to show the U-matrix and the component
Figure 3. Data set preprocessing tool.
planes of the SOM:

som_show(sM);

The U-matrix visualizes distances between neighboring

map units, and thus shows the cluster structure of the map:
high values of the U-matrix indicate a cluster border,
uniform areas of low values indicate clusters themselves.
Each component plane shows the values of one variable in
each map unit. On top of these visualizations, additional
information can be shown: labels, data histograms and
trajectories.
With function som_vis much more advanced
visualizations are possible. The function is based on the
idea that the visualization of a data set simply consists of a
set of objects, each with a unique position, color and
shape. In addition, connections between objects, for
example neighborhood relations, can be shown using
lines. With som_vis the user is able to assign arbitrary
values to each of these properties. For example, x-, y-, and
z-coordinates, object size and color can each stand for one
variable, thus enabling the simultaneous visualization of
five variables. The different options are:
- the position of an object can be 2- or 3-dimensional
- the color of an object can be freely selected from
the RGB cube, although typically indexed color is
used
Figure 4. SOM initialization and training tool. - the shape of an object can be any of the Matlab
plot markers ('.','+', etc.), a pie chart, a bar
 chart, a plot or even an arbitrarily shaped polygon, %% make the data
typically a rectangle or hexagon sD = som_read_data('iris.data');
- lines between objects can have arbitrary color, sD = som_normalize(sD,'var');
width and any of the Matlab line modes, e.g. '-'
- in addition to the objects, associated labels can be %% make the SOM
shown sM = som_make(sD,'munits',30);
For quantitative analysis of the SOM there are at the sM = som_autolabel(sM,sD,'vote');
moment only a few tools. The function som_quality
supplies two quality measures for SOM: average %% basic visualization
quantization error and topographic error. However, using som_show(sM,’umat’,’all’,’comp’,1:4,...
low level functions, like som_neighborhood, ’empty’,’Labels’,’norm’,’d’);
som_addlabels(sM,1,6);
som_bmus and som_unit_dists, it is easy to
implement new analysis functions. Much research is being
From the U-matrix it is easy to see that the top three
done in this area, and many new functions for the analysis
rows of the SOM form a very clear cluster. By looking at
will be added to the Toolbox in the future, for example
the labels, it is immediately seen that this corresponds to
tools for clustering and analysis of the properties of the
the Setosa subspecies. The two other subspecies
clusters. Also new visualization functions for making
Versicolor and Virginica form the other cluster. The U-
projections and specific visualization tasks will be added
matrix shows no clear separation between them, but from
to the Toolbox.
the labels it seems that they correspond to two different
4.6. Example parts of the cluster. From the component planes it can be
seen that the petal length and petal width are very closely
Here is a simple example of the usage of the Toolbox to related to each other. Also some correlation exists between
make and visualize a SOM of a data set. As the example them and sepal length. The Setosa subspecies exhibits
data, the well-known Iris data set is used [5]. This data set small petals and short but wide sepals. The separating
consists of four measurements from 150 Iris flowers: 50 factor between Versicolor and Virginica is that the latter
Iris-setosa, 50 Iris-versicolor and 50 Iris-virginica. The has bigger leaves.
measurements are length and width of sepal and petal
U−matrix sepallength sepalwidth
leaves. The data is in an ASCII file, the first few lines of 1.6 7.09
3.7
which are shown below. The first line contains the names 1.4 6.67
3.49
1.2
of the variables. Each of the following lines gives one 1
6.26
3.27
data sample beginning with numerical variables and 0.8 5.84 3.05

followed by labels. 0.6 5.43 2.84

0.4 5.02 2.62
0.2
#n sepallen sepalwid petallen petalwid d d

5.1 3.5 1.4 0.2 setosa

petallength petalwidth Labels
4.9 3.0 1.4 0.2 setosa se se se se se se
5.52 1.96 se se se se se se
... se se se se se
se se se
4.64 1.58 ve
ve ve ve ve ve ve
ve ve ve ve ve vi
The data set is loaded into Matlab and normalized. 3.76 1.2 ve ve ve ve ve ve
ve ve ve ve
Before normalization, an initial statistical look of the data 2.88 0.817 vi ve ve ve
vi ve vi vi vi vi
vi vi vi vi
set would be in order, for example using variable-wise 1.99 0.436 vi
vi
vi vi vi
vi vi vi vi
histograms. This information would provide an initial idea d d

of what the data is about, and would indicate how the Map: SOM 06−Sep−1999, Size: 14 6

variables should be preprocessed. In this example, the
variance normalization is used. After the data set is ready,
a SOM is trained. Since the data set had labels, the map is
also labeled using som_autolabel. After this, the
SOM is visualized using som_show. The U-matrix is
shown along with all four component planes. Also the
labels of each map unit are shown on an empty grid using
som_addlabels. The values of components are
denormalized so that the values shown on the colorbar are
in the original value range. The visualizations are shown
in Figure 5.
Figure 5. Visualization of the SOM of Iris data. U-
matrix on top left, then component planes, and map unit
labels on bottom right. The six figures are linked by
position: in each figure, the hexagon in a certain position
corresponds to the same map unit. In the U-matrix,
additional hexagons exist between all pairs of neighboring
map units. For example, the map unit in top left corner has
low values for sepal length, petal length and width, and
relatively high value for sepal width. The label associated
with the map unit is 'se' (Setosa) and from the U-matrix it
can be seen that the unit is very close to its neighbors.
 Component planes are very convenient when one has to 5. Conclusions
visualize a lot of information at once. However, when only
a few variables are of interest scatter plots are much more In this paper, the SOM Toolbox has been shortly
efficient. Figures 6 and 7 show two scatter plots made introduced. The SOM is an excellent tool in the
using the som_grid function. Figure 6 shows the PCA- visualization of high dimensional data [6]. As such it is
projection of both data and the map grid, and Figure 7 most suitable for data understanding phase of the
visualizes all four variables of the SOM plus the knowledge discovery process, although it can be used for
subspecies information using three coordinates, marker data preparation, modeling and classification as well.
size and marker color. In future work, our research will concentrate on the
3 quantitative analysis of SOM mappings, especially
analysis of clusters and their properties. New functions
and graphical user interface tools will be added to the
2
Toolbox to increase its usefulness in data mining. Also
se vi outside contributions to the Toolbox are welcome.
se se
1 se vi vi It is our hope that the SOM Toolbox promotes the
sese vi
se ve vi
vi vi vi utilization of SOM algorithm – in research as well as in
sese ve vi vi vi vi
se ve viveve industry – by making its best features more readily
ve vi
0 sese se veve accessible.
se veve vevi
se ve vi vi
sese veve ve vi
se ve ve
se veve vivi
Acknowledgements
−1 ve
se ve ve ve
veve
This work has been partially carried out in ‘Adaptive
ve and Intelligent Systems Applications’ technology program
−2
of Technology Development Center of Finland, and the
EU financed Brite/Euram project ‘Application of Neural
−3 Network Based Models for Optimization of the Rolling
−3 −2 −1 0 1 2 3 4
Process’ (NEUROLL). We would like to thank Mr. Mika
Figure 6. Projection of the IRIS data set to the Pollari for implementing the initialization and training
subspace spanned by its two eigenvectors with greatest GUI.
eigenvalues. The three subspecies have been plotted using
different markers: IRU6HWRVDx for Versicolor and ¸IRU
References
Virginica. The SOM grid has been projected to the same [1] Kohonen T. Self-Organizing Maps. Springer, Berlin, 1995.
subspace. Neighboring map units connected with lines. [2] Vesanto J., Alhoniemi E., Himberg J., Kiviluoto K.,
Labels associated with map units are also shown. Parviainen J. Self-Organizing Map for Data Mining in
MATLAB: the SOM Toolbox. Simulation News Europe
1999;25:54.
[3] Kohonen T., Hynninen J., Kangas J., Laaksonen J.
SOM_PAK: The Self-Organizing Map Program Package,
7 Technical Report A31, Helsinki University of Technology,
6
1996, http://www.cis.hut.fi/nnrc/nnrc-programs.html
[4] Pyle D. Data Preparation for Data Mining. Morgan
5 Kaufman Publishers, San Francisco, 1999.
4
[5] Anderson E. The Irises of the Gaspe Peninsula. Bull.
American Iris Society; 1935;59:2-5.
3 [6] Vesanto J. SOM-Based Visualization Methods. Intelligent
2
Data Analysis 1999;3:111-126.

1
4
Address for correspondence.
7.5
3.5
7
3 6.5 Juha Vesanto
6
2.5 5.5
Helsinki University of Technology
2 4.5
5 P.O.Box 5400, FIN-02015 HUT, Finland
Juha.Vesanto@hut.fi
Figure 7. The four variables and the subspecies http://www.cis.hut.fi/projects/somtoolbox
information from the SOM. Three coordinates and marker
size show the four variables. Marker color gives
subspecies: black for Setosa, dark gray for Versicolor and
light gray for Virginica.